Duke Humanoid V1

There is no price and no supplier. The Duke Humanoid is a design to build from, like the SO-100 and OpenArm, not a product to order. Duke has not published a build cost: the paper discusses the cost of geared motors as a design constraint but gives no bill-of-materials total, so the accurate entry is open hardware, build cost not published.

The Duke Humanoid is a 10-degree-of-freedom bipedal robot released as a complete open platform: hardware documentation, CAD, and the reinforcement learning stack that controls it. It stands about 1 meter from shoulder to foot and weighs 30 kg, roughly the proportions of an eight-year-old child. Duke chose that size deliberately: large enough to reach human-height surfaces for later manipulation work, small enough to stay affordable given the cost of geared motors. It sits alongside the K-Scale Zeroth Bot and the LeRobot Humanoid as an open-source bipedal platform built for research rather than sale.

All ten degrees of freedom are in the legs, five per side: three at the hip (rotation, flexion/extension, abduction/adduction), one at the knee, and one ankle limited to plantar flexion and dorsiflexion. The ankle was constrained to a single axis to reduce mass at the end of the limb. Leg proportions and joint ranges follow human anatomy, and the hip's center of mass is aligned in the frontal plane so the robot can balance and walk with straight knees.

The research contribution is the control rather than the mechanics. Most walking robots hold every joint under active control throughout the gait, while humans and animals let the leg swing freely through much of it. The team trained an end-to-end reinforcement learning policy that alternates between active and passive control, reaching a 50% improvement in cost of transport in simulation and 31% on the physical robot at low walking speed. Training runs in Isaac Gym and Isaac Lab.

The paper is arXiv:2409.19795, The Duke Humanoid: Design and Control For Energy Efficient Bipedal Locomotion Using Passive Dynamics (Xia, Li, Lee, Scutari, Chen). The code is in the DukeHumanoidv1 repository, which covers the RL simulation and hardware control and has 72 stars. The same lab also publishes Text2Robot (generating robot designs from text prompts and optimizing them with evolutionary algorithms), ClutterGen, SonicFly and MARBLE.